Internal representation of two-dimensional shape

Area of Science:

• Cognitive Psychology
• Neuroscience
• Computational Neuroscience

Background:

• The internal representation of shapes in the brain remains underexplored in psychological literature.
• Previous experiments studied the psychological space of shapes but lacked a unified explanation for internal representation formation.

Purpose of the Study:

• To investigate the characteristics and formation of internal shape representations in the brain.
• To unify experimental findings on shape perception through a novel theoretical framework.

Main Methods:

• Reexamination of similarity data for alphabetic characters and random-dot patterns using multivariate analysis.
• Training three-layer neural networks for pattern categorization and identity transformation tasks.
• Analysis of hidden unit activation patterns in neural networks to understand internal representations.

Main Results:

• Multivariate analysis indicated that patterns are represented by a combination of global features.
• Neural networks trained for categorization did not yield representations matching psychological data.
• Networks trained for identity transformation acquired representations similar to those suggested by psychological experiments.

Conclusions:

• The human brain likely represents two-dimensional shapes through a process analogous to principal-component analysis.
• This principal-component analysis-like process involves projecting input images onto a lower-dimensional space.
• The proposed representation mechanism is consistent with findings from neurophysiological studies of higher visual areas.